{"id":10614,"date":"2017-11-13T06:01:41","date_gmt":"2017-11-13T10:01:41","guid":{"rendered":"http:\/\/whamit.mit.edu\/?p=10614"},"modified":"2017-11-12T18:48:39","modified_gmt":"2017-11-12T22:48:39","slug":"complang-1116-david-alvarez-melis-mit-csail","status":"publish","type":"post","link":"https:\/\/ling-phil.mit.edu\/whamit\/2017\/11\/13\/complang-1116-david-alvarez-melis-mit-csail\/","title":{"rendered":"CompLang 11\/16 - David Alvarez Melis (MIT CSAIL)"},"content":{"rendered":"<p><strong>Speaker:<\/strong>\u00a0<a href=\"http:\/\/people.csail.mit.edu\/davidam\/\">David Alvarez Melis<\/a> (MIT CSAIL)<br clear='none'\/>\n <strong>Title:<\/strong> Interpretability for black-box sequence-to-sequence models<br clear='none'\/>\n <strong>Date\/Time:<\/strong> Thursday, November 16th, 5:00-6:30pm<br clear='none'\/>\n <strong>Location:<\/strong> 46-5165<br clear='none'\/>\n <strong>Abstract:<\/strong><\/p>\n<blockquote>\n<div class=\"b5 xJNT8d\">\n<div class=\"uyb8Gf\">\n<div class=\"F3hlO\">\n<div dir=\"ltr\">Most current state-of-the-art models for sequence-to-sequence NLP tasks have complex architectures and millions &#8212;if not billions&#8212;of parameters, making them practically\u00a0black-box systems. Such lack of transparency can limit their applicability to certain domains and can hamper our ability to diagnose and correct their flaws. Popular black-box\u00a0interpretability approaches are inapplicable to this context since they assume scalar (or categorial) outputs. In this work, we propose a model to interpret the predictions of any black-box\u00a0structured input-structured output model around a specific input-output pair. Our method returns an &#8220;explanation&#8221; consisting of groups of input-output tokens that are causally related.\u00a0These dependencies are inferred by querying the black-box model with perturbed inputs, generating a graph over tokens from the responses, and solving a partitioning problem to select\u00a0the most relevant components. We focus the general approach on sequence-to-sequence problems, adopting a variational autoencoder to yield meaningful input perturbations. We test\u00a0our method across several NLP sequence generation tasks.<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Speaker:\u00a0David Alvarez Melis (MIT CSAIL) Title: Interpretability for black-box sequence-to-sequence models Date\/Time: Thursday, November 16th, 5:00-6:30pm Location: 46-5165 Abstract: Most current state-of-the-art models for sequence-to-sequence NLP tasks have complex architectures and millions &#8212;if not billions&#8212;of parameters, making them practically\u00a0black-box systems. Such lack of transparency can limit their applicability to certain domains and can hamper our [&hellip;]<\/p>\n","protected":false},"author":18,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-10614","post","type-post","status-publish","format-standard","hentry","category-talks"],"_links":{"self":[{"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/posts\/10614","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/users\/18"}],"replies":[{"embeddable":true,"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/comments?post=10614"}],"version-history":[{"count":2,"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/posts\/10614\/revisions"}],"predecessor-version":[{"id":10624,"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/posts\/10614\/revisions\/10624"}],"wp:attachment":[{"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/media?parent=10614"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/categories?post=10614"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ling-phil.mit.edu\/whamit\/wp-json\/wp\/v2\/tags?post=10614"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}